WAIVIER is a Checkpoint Merge of some of my favorite Checkpoints!
Please check the 'About this version' section for more information about what models were used in the merge, at what weights, and what might have changed between versions.
I originally made this just for myself, intending to keep it as a private merge, but as I uploaded more images and posts without having a Checkpoint to link them as resources used, I decided to just upload it so others can use it too.
Recommended Settings:
CFG: 3
Steps: 25 - 40
Sampler: DPM++ 2M SDE
Scheduler: Karras
Clip Skip: 2
VAE Baked-in
I will always recommend using HiRes Fix with all generations to really sharpen things up, but I generate locally so I don't need to budget my buzz for it. It works without the HiRes upscaling process, but the final result will always be cleaner and sharper with it.
This is a learning experience for me, so please have patience with me. I welcome criticism, feedback, requests, and suggestions!
All Rights and permissions are as you'd expect when it comes to a merge. Please generate responsibly.
Description
v2.0 Release. (Full Release)
VAE Baked-In
Clip Skip 2
Merged locally using SuperMerger extension for Stable Diffusion Forge. A handful of personal style LoRAs, private Checkpoint merges, and additional poses and concepts were included in the merging process in order to get the visual style as I like it.
Checkpoint is fp16, pruned, and has VAE baked-in. I recommend Clip Skip 2 as that's what I use and that's what it was merged with.
My Recommended Settings:
CFG: 3.0 - 7.0
Steps: 25 - 40
Sampler: DPM++ 2M SDE
Scheduler: Karras
Clip Skip: 2
It's recommended to always Hires Upscale all images generated, as to ensure quality outputs. This is, of course, a personal choice though.
Please keep in mind these settings tend to work best for me. That does not mean they will work best for you. Experiment and have fun.
This is a continuation of my existing v1 Checkpoint, with some slight alterations to concepts, poses, and visual styles. The base style is still the same as v1, with relatively minimal changes between iterations, with a heavier focus on being more friendly to furry themes and newer characters.
The base model used was my WAIVIER v1 Release, with some private merges and loras used to help train more poses and concepts. Understanding of characters should also be better, and any continued blind-spots in the model can be focused on and improved in future versions.
I was originally planning on holding off on releasing a version 2, but user feedback showed there was interest in an updated model, so wanted to give it a try. There's not much of a difference between v1 and v2 stylistically, so you should still be able to use mostly the same prompt tags to get solid results.
FAQ
Comments (24)
This looks so cool. :D I can't wait to try it out!
Gorgeous, intelligent model.
The new version is gorgeous!
Beautiful model
Need a new set of picture quality control and style prompts. The picture quality control prompts from the old 1.0 version don’t seem to work well when I set the high-res fix to a very high value with version 2.0. The speed of the generation seems faster compare to the v1.0.
>Need a new set of picture quality control and style prompts.
Yeah, working on it. Most prompts from 1.0 should work alright, but some prompt adherence has been changed between versions, which could explain things. v1.0 has had half a year of me putting out public prompts that have passed a quality-assurance test by me, so v2.0 will take a bit more time to find the perfect prompt formatting to give the best outputs possible. Experiment, try new things, and post your successes, would be really cool to see!
>The picture quality control prompts from the old 1.0 version don’t seem to work well when I set the high-res fix to a very high value with version 2.0.
Focusing heavily on Danbooru tags helps, also tags from rule34 should also be useful (hopefully not under-trained, but let me know how you find it.)
The visual style is mostly the same between v1.0 and v2.0 but the prompting may need some further adjusting. I always recommend using HiRes upscaling on all images you plan on posting, just because it really helps the final output, but finding the best settings is very much to-taste.
For those interested, my typical HiRes settings are as follows:
Upscaler - RealESRGAN_x4plus_anime_6B (I also sometimes use 4x-UltraSharpV2, but for most of my anime stuff I use the former.)
Hires steps - 20, Denoising strength - 0.25, Upscale by - x2, Hires Sampling method - Euler a, and Hires CFG Scale - 5.
I've tried higher denoising strengths, but it tends to cause issues where characters have extra nipples, or too many abs, etc etc. Experiment yourself and let me know your results! :D
>The speed of the generation seems faster compare to the v1.0.
Happy to hear!
Also, I've put a bid on both v1.0 and v2.0 for this following week, so hopefully both versions will be available for generation on-site, so as to make it easier to experiment and test results.
@MysticFrostwind @Tony_Sensei
First off, great model. really unique stuff.
V2.0 for me was more sensitive towards prompts. So you can lower your recommended CFG down from 5-7 to more like 3-5.
Epsecially if you are going to recommend that your users choose any other scheduler/sampler besides Euler A (normal)
Best results IMO:
Steps: 20 - 25
Denoise: 0.7
Resolutions to use: 1024x1024 (square), 832x1216(Portrait), 1216x832(Landscape)
HiRes Steps: (steps * 0.5)
Denoise: 0.2 ~ 0.4 (If you have the option to select)
HiRes Upscale: 1.5
CFG: 4
Sampler/Scheduler: Euler A (Normal OR simple) for all
Any of 4x upscalers are good ( R-ESRGAN, ESRGAN, or Remacri) are fine to use
@MysticFrostwind It is the negative prompt I just figure out. Certain key words need to be remove such as crisp and correct position. I also adding distortion nipples and distortion abs into the negative prompt as the broken parts usually appear around these position. Now 8/10 chance will get perfect result even when the high fix stick in 5.0.
@Tony_Sensei Great testing!
peak. i love it! it has become my 2nd to-go model, specially for lora trials. it just /works/, yk. one of myfav models on this site, uwa ♡🥹
Having fun working iwth the model - was wondering if there's some wayu to speed it up. ~80 sec time to generate on a 5060ti is rough
It definitely shouldn't be taking that long, especially on a 50 series RTX. I've generated locally even with a 3060Ti which only has 8GB of VRAM with minimal gen times, so there's definitely something going wrong if it's taking 80 seconds per gen for you.
My experience is with SD Forge mostly, so can't give much help if you're using ComfyUI. If I'm able to help, though, let me know!
From my experience using only 8GB VRAM a single image gen at 832x1216 before Hires takes about 20 seconds or so, with Hires adding another 20 seconds or so afterwards. If using Adetailer for face+eyes the time to generated may increase, but all in all a 5060Ti should outperform my gen times by a fair bit. Weird, very weird.
I'll do some research to see if there's anything more on my end that I can do to help speed things up for you.
It's funny how there are models (like this one) that seem to have been created exclusively for gay. I scrolled through all the user examples, and literally only two images show a woman.
The model is based on Illustrious so we already know it's more than capable of hetero themes haha. I'm gay, though, so I tend to focus my attention toward content that I prefer and I test prompts with that in mind.
More than welcome to post any generative work you make, straight or gay. So far I've not needed to moderate the Gallery section, and I have no intentions of doing so. :D
90% of models on this site are primarily focused and tested with women tags, so wanted to try my hand at a male focus.
@MysticFrostwind This isn't nitpicking or an insult. I always browse other people's work under “Lora” or “Models” to figure out if I need it or not. And if I do, I copy it to a folder with my selected works so I can remember why I added a particular program.
The model is good. I was just surprised that the entire model features only one type of content. This is the first time I’ve seen that.
@Twostep No worries man, I took no offense from it to begin with. Glad you like the model, I'm typically just long-winded and a tad bit awkward. >.>
This is great good job
Thanks! :)
Love this model so much. Are you going to update this to ANIMA?
It's not impossible, I just don't have anything atm that can run and generate ANIMA realistically, since my stuff is generated/merged/trained locally.
It's definitely been on my radar, though.
who is this?
What do you mean?
@MysticFrostwind mean this model?
@MysticFrostwind and fine tune merge